Stas Bekman wrote:
Philippe M. Chiasson wrote:
Geoffrey Young wrote:
The keyword is the "not immediately" -- that solution will make Apache::SizeLimit and its clones unreliable. Since you will never be able to bracket the memory usage again.
That's incorrect. You will be able to bracket memory usage. In a prefork MPM, calling $r->child_terminate() will kill off that child as soon as the current request is done with, after running cleanups et all.
OK, so what's the difference from calling exit?
There are 2 major differences. First, you will get proper cleanup and all pool cleanups will be run. Second, you will get a child that gets to cleanly shutdown, however the child chooses to do that.
For a threaded MPM like prefork, we currently have no valid solution to
implement child_terminate() in order to reclaim memory. This way would at
least provide a partial solution. If a child grows to fat, calling $r->child_terminate()
would guarantee that that child will not serve new requests (keepalive issues exluded)
and kill itself off as soon as it can.
But it's very dangerous. Assume that you have N processes with M threads. If all processes will be scheduled to be killed, you will have a long long time w/o the server accepting any new requests.
No, no, only the child process will do that, the parent, and all the other siblings will keep on hapilly serving requests.
Ofcourse the apache 1.3 api was calling child_terminate at the end of the request, so it wasn't immediate either, but it was guaranteed to be called asap. I believe with signals it's totally unreliable.
No, it's about as reliable as apachectl restart/gracefull will kill off the childs.
but we aren't trying to call 'graceful restart'
No, we are making the child process believe it's parent is about to do just that.
in 1.3 ap_child_terminate just sets MaxRequestsPerChild to 0, leaving apache
to reap the child at the end of the current request.
Yup, I looked at that, but in 2.0, at the beginning of the lifetime of a child,
a private counter is set to MaxRequestsPerChild and decremented for each request.
Child termination is caused when that number reaches zero. So, changing MaxRequestPerChild
will not work that way anymore.
So any chance we can get hold of that counter?
Nope, most MPMs I looked at implement that as a local var.
-- -------------------------------------------------------------------------------- Philippe M. Chiasson m/gozer\@(apache|cpan|ectoplasm)\.org/ GPG KeyID : 88C3A5A5 http://gozer.ectoplasm.org/ F9BF E0C2 480E 7680 1AE5 3631 CB32 A107 88C3A5A5
signature.asc
Description: OpenPGP digital signature
